I think that even if it would be trained on a large set of actual CSAM data, it makes a very practical difference that the "marginal cost" of new images doesn't

I think that even if it would be trained on a large set of actual CSAM data, it makes a very practical difference that the "marginal cost" of new images doesn't imply more abuse. E.g. I have a local example in mind after a local pedophile was arrested in a bust of an international CSAM ring, the story goes that he was consuming CSAM, and joining that "better" ring of CSAM content required providing "original content" (that's one way they protected against law enforcement infiltration) and allegedly that need for that was the reason he went from watching CSAM to actually molesting several kids to film it. If we can remove this kind of motivation, that's IMHO a good thing for society.